Abstract
Although image inpainting is now an effective image editing technique, limited work has been done for inpainting forensics. The main drawbacks of the conventional inpainting forensics methods lie in the difficulties on inpainting feature extraction and the very high computational cost. In this paper, we propose a novel approach based on a convolutional neural network (CNN) to detect patch-based inpainting operation. Specifically, the CNN is built following the encoder–decoder network structure, which allows us to predict the inpainting probability for each pixel in an image. To guide the CNN to automatically learn the inpainting features, a label matrix is generated for the CNN training by assigning a class label for each pixel of an image, and the designed weighted cross-entropy serves as the loss function. They further help to strongly supervise the CNN to capture the manipulation information rather than the image content features. By the established CNN, inpainting forensics does not need to consider feature extraction and classifier design, and use any postprocessing as in conventional forensics methods. They are combined into the unique framework and optimized simultaneously. Experimental results show that the proposed method achieves superior performance in terms of true positive rate, false positive rate and the running time, as compared with state-of-the-art methods for inpainting forensics, and is very robust against JPEG compression and scaling manipulations.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have